CAPABILITIES & METHODS · 15 OCTOBER 2026
What works here.
What comes next.
Arc Corpus Toolkit is an independent browser corpus workspace. Its available tools, methods and current limits are listed below.
Open Arc Corpus Toolkit →Feature coverage
| Research workflow | Arc Corpus Toolkit implementation | Scope |
|---|---|---|
| Concordance, KWIC and dispersion | Words, phrases, wildcards; simple regex; token positions; CSV/JSON exports; sorting and subcorpus filters | Operational, browser scale. Up to 5,000 retained result rows. |
| Corpus Query Language | Sequential [word="…"], [lemma="…"] and [tag="…"] tests | Subset only. No full CQL quantifiers, structures, Boolean expressions or regex groups. |
| Word lists, n-grams and clouds | Surface frequencies, supplied lemmas/tags, 2–5-grams, document frequency and visual cloud | Operational. English stopword list is small and explicit. |
| Collocations and grammatical word profiles | Window associations; separate dependency-based word profiles from CoNLL-U | Plain text is not automatically tagged. Profiles use imported HEAD/DEPREL/LEMMA/UPOS. |
| Grammatical profile comparison / word comparison | Side-by-side window collocation counts and score differences | Partial analogue; grammatical profile comparison and bilingual sketches are not implemented. |
| Thesaurus | Cosine similarity over local context counts | Limited to 400 frequent candidates. Related contexts do not establish synonymy. |
| Keyword and term extraction | Target/reference log likelihood and smoothed log₂ ratios; n-gram term candidates | Term candidates require review. No complete language-specific term grammar. |
| Parallel concordance | Two-column aligned TSV corpus and source-side search | Operational for manually aligned pairs. No automatic alignment, translation or TMX/XLIFF ingestion. |
| Bilingual term extraction | Unigram source–target co-occurrence candidates and Dice scores | Experimental candidate generation. Not full bilingual terminology extraction or full multilingual terminology extraction. |
| Trends, timeline and text types | Per-year and per-category normalized query frequencies; charts | Operational with user metadata. No automated neologism discovery or significance models. |
| Word sense induction | Experimental context clusters from three seed examples and lexical overlap | Not trained word-sense models or semantic disambiguation. |
| Dictionary examples / a trained example-ranking model | Length-ranked example finder | No a trained example-ranking model scoring model, dictionary editor or TickBox Lexicography. |
| Corpus building and subcorpora | Local files, pasted text, annotations, aligned TSV, metadata and file selection | No automated web crawling or hosted corpus compilation. |
| Reading and separate preparation workspaces | File reading stays here. Corpus Cleaner, File Converter and Corpus Merger are separate workspaces on the Toolkits page. | Source data stays unchanged. Reference/citation removal uses English heuristics and needs review. No OCR or original layout preservation. |
| Statistics and charts | Document size, estimated sentence length, type/token ratio; bars, lines and donuts for numeric results; SVG download | Up to 20 chart rows in table order. Line order is not necessarily time; donut uses positive values. Tables retain exact results. |
| English grammar review | Local agreement, selected articles, repeated words, pronoun capitalization, punctuation and countability checks; evidence with line/offset | Heuristics, not a complete grammar or spelling model. No automatic correction or validated proficiency score. Review false positives. |
| Duplicate paragraphs | Exact paragraph matches within/across selected documents after whitespace normalization | Available in the separate Corpus Cleaner. At least 30 characters, case-sensitive. No semantic plagiarism detection. Removal is opt-in. |
| Downloads, history and favorites | TXT, SVG charts, CSV, JSON and local workspace backups; saved views and optional browser remembering | Local only. No shared corpus links, team workspaces or institutional SSO. |
| Ready-made corpora and multilingual annotation | Illustrative sample data and imported annotations | No licensed reference-corpus library or automatic multilingual POS/lemma/dependency models. |
| Hosted APIs, large corpora, crawling, team accounts and CAT integration | Not available in this static browser release | Require separate server infrastructure, resources and integration work. No release date is promised. |
Data and performance
All corpus processing stays in your browser. Analyses run in a dedicated worker and time out after 30 seconds (60 seconds for corpus indexing) rather than blocking the page indefinitely. Desktop ceiling: 500 files per corpus; 100 million characters and 3 million tokens across target + reference; 50 million characters / 250 MB per file. Mobile/low-memory: 250 files, 20 million characters / 1 million tokens total, 10 million characters / 60 MB per file. Selectable PDFs: 10,000 pages. Aligned rows: 10,000. Device RAM can reduce practical capacity.
Corpus backups contain your text. Sharing or downloading a file is your own choice. Browser storage may be cleared by the device; keep an exported backup. PDF and DOCX readers load from this site. Scanned PDFs need OCR before import. No offline PWA is enabled.